Regensburg 2022 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

O: Fachverband Oberflächenphysik

O 67: Frontiers of Electronic Structure Theory: Focus on Artificial Intelligence Applied to Real Materials 3

O 67.1: Vortrag

Donnerstag, 8. September 2022, 10:30–10:45, S054

Quantile Random Forest Model for Extrapolation to the Complete Basis Set Limit in Density Functional Theory Calculations — •Daniel Speckhard¹, Christian Carbogno², Sven Lubeck², Luca Ghiringhelli², Matthias Scheffler^{2, 1}, and Claudia Draxl^{1, 2} — ¹Humboldt-Universität zu Berlin, Physics Department and IRIS Adlershof, Berlin, Germany — ²The NOMAD Laboratory at the FHI-MPG and HU, Berlin, Germany

The precision of density-functional theory (DFT) calculations depends on a variety of computational parameters, the most critical being the basis-set size. With an infinitely large basis set, i.e., in the limit of a complete basis set (CBS), the result of the calculation is as precise as possible for the chosen exchange-correlation functional. Our aim in this work is to find a model that can extrapolate the result of an imprecise DFT calculation to the CBS limit. As a starting point, we use a dataset of 63 binary solids investigated with various basis-set sizes [1] with two all-electron DFT codes, exciting and FHI-aims, which use very different types of basis sets. A quantile random forest model is used to estimate the deviation of the total energy with respect to fully converged calculations as a function of the basis set size. The non-linear random forest model outperforms a previous approach that used a linear model. The quantile random forest model presented also provides prediction intervals which give the user an idea of the model’s uncertainty.

[1] C. Carbogno et al., npj Comput. Mater. 8, 69 (2022).